2 ms·

Another benefit of whiteboard interviews that I often see left out is how well they scale. You can train a lot of engineers in how to conduct these interviews

Another benefit of whiteboard interviews that I often see left out is how well they scale. You can train a lot of engineers in how to conduct these interviews fairly quickly and you tell them what information to focus on and how to document it. You can then pass this information to a hiring committee that has a lot of experience in evaluating candidates. Having sat in on some hiring committee meetings I can see how little weight is given the algorithm if the candidate has a good reference or highly relevant experience. The issue with the interview alternatives that most people suggest is that they would require your most senior engineers to spend a really really large portion of their day on hiring which can burn people out quickly.

dheera 6y ago
My main issue with whiteboard interviews is that they often ask you to code something that is completely unrelated to the job, and is not representative of your normal work efficiency at a task that is within your area of expertise. For example, I've had several machine learning engineer interviews in which I was asked to write some variation of NMS (non-max suppression) or gradient descent on the whiteboard. What machine learning engineer writes that kind of code anymore? I can write it, yes I can fumble (a lot) and figure out how to optimize it, but in reality these are the type of algorithms that you almost always use some library that someone has already written for you for that kind of stuff. They spend the entire hour nitpicking at your implementation of NMS, base their entire impression of you on that implementation of NMS, and never in your entire interview ask you anything about, say, what the state-of-the-art object detectors are, what their breakthroughs are, how their networks are roughly structured, what you might think about doing to improve on them, what you could do if you only had limited annotation data, and other real machine learning questions.
- blamestross 6y ago
  Actually evaluating your knowledge is hard. Picking something to be a rough litmus test and testing that, despite it not being directly related to the job seems reasonable. It has been 3 years since I interviewed somebody for a ML position. Mostly I focused on knowledge of model execution. Most ML libraries were horrible for use in production at the time and I was more worried that a candidate wouldn't be able to productionize a good model than I was worried they wouldn't be able to build that model.
ThrowawayR2 6y ago
> "Having sat in on some hiring committee meetings I can see how little weight is given the algorithm if the candidate has a good reference or highly relevant experience." I've sat in on a fair number of hiring committee meetings over the years and have only seen the opposite; poor performance on the testing is a disqualifier no matter how good their references or credentials are. Even the FAANGs make hiring mistakes so credentials don't mean much.

dheera 6y ago

My main issue with whiteboard interviews is that they often ask you to code something that is completely unrelated to the job, and is not representative of your normal work efficiency at a task that is within your area of expertise. For example, I've had several machine learning engineer interviews in which I was asked to write some variation of NMS (non-max suppression) or gradient descent on the whiteboard. What machine learning engineer writes that kind of code anymore? I can write it, yes I can fumble (a lot) and figure out how to optimize it, but in reality these are the type of algorithms that you almost always use some library that someone has already written for you for that kind of stuff. They spend the entire hour nitpicking at your implementation of NMS, base their entire impression of you on that implementation of NMS, and never in your entire interview ask you anything about, say, what the state-of-the-art object detectors are, what their breakthroughs are, how their networks are roughly structured, what you might think about doing to improve on them, what you could do if you only had limited annotation data, and other real machine learning questions.

blamestross 6y ago
Actually evaluating your knowledge is hard. Picking something to be a rough litmus test and testing that, despite it not being directly related to the job seems reasonable. It has been 3 years since I interviewed somebody for a ML position. Mostly I focused on knowledge of model execution. Most ML libraries were horrible for use in production at the time and I was more worried that a candidate wouldn't be able to productionize a good model than I was worried they wouldn't be able to build that model.

ThrowawayR2 6y ago

> "Having sat in on some hiring committee meetings I can see how little weight is given the algorithm if the candidate has a good reference or highly relevant experience." I've sat in on a fair number of hiring committee meetings over the years and have only seen the opposite; poor performance on the testing is a disqualifier no matter how good their references or credentials are. Even the FAANGs make hiring mistakes so credentials don't mean much.